His data showed the AI was hurting veterans' claims - the agency approved his blog, then let him go
A Presidential Innovation Fellow told the Department of Veterans Affairs (VA) its AI tool was broken. The VA fired him.
The US Court of Appeals for the Federal Circuit reversed the Merit Systems Protection Board's dismissal of the worker's whistleblower appeal on September 22, finding he had made sufficient allegations of gross mismanagement to establish jurisdiction.
The worker joined the VA in October 2020 with a specific brief: help cut the time veterans wait for disability benefits decisions. The agency's solution was an AI tool called Claims Attribute Application Programming Interface (CAAPI), pitched as shaving three to five days off decision times and saving $1.5 million in direct labor costs.
The worker tested it. CAAPI "was giving wrong predictions." His initial dataset showed the tool's claimed gains "were not tethered to the truth," and he flagged that to his supervisor in March 2021. A larger analysis of 716,000 claims told the same story - CAAPI "was slowing down the processing of claims by 5 days on average and 9 weeks in the case of complicated submissions."
He published a blog post about the problems in April 2021. The VA approved it.
The agency's next move was to repackage CAAPI into a new version - "without any modification to the classifier that underpins the API" - and put it back into production. The worker escalated to the chief technology officer in May. By July, the tool was pulled because "the model underlying the Claims Attributes API likely does not reduce Veteran benefit decision wait times."
The day before that, the chief technology officer recommended ending the worker's contract. He was removed July 7, 2021. His performance rating: "Fully Successful."
The MSPB threw out his whistleblower appeal at the jurisdictional stage. The Federal Circuit reversed, holding his disclosures were "detailed and facially well-supported" - not "vague, conclusory or facially insufficient." With the VA processing over 1.5 million claims a year, deploying an unmodified tool on live claims was plausibly serious error "not debatable among reasonable people."
A concurring opinion cautioned the ruling should not discourage experimentation to solve the VA's claims backlog - but noted the agency "appears to have been unresponsive to reasonable attempts" to raise concerns.
For HR teams rolling out AI tools, the takeaway lands close to home - ignoring data-backed objections and then removing the person who raised them is exactly the sequence whistleblower statutes are built to catch.
